Photographic Composite Detection Using Scene Perspective Constraints

The prevalence of image acquisition devices and powerful processing software have given rise to large amounts of doctored images without obvious traces,triggering a great demand for automatic forgery detection algorithms that can identify the trustworthiness of a candidate image.Passive image forensics has attracted great research interests since it does not require any auxiliary data such as watermarks or signatures and directly identifies the authority of images based on inhere statistical properties.Two algorithms based on scene constraints for photographic composite detection have been proposed in this thesis.Scene constraints means the inherently existed geometric properties of scene corresponding to the image,such as the relative position and the height ratio.The innovations and contribution of the paper are as follows:Firstly,a passive forensics approach based on the property of vanishing point is proposed to detect composition.Parallel segments in the scene intersect at a unique point within or outside the image captured by pinhole camera,namely the vanishing point.However,noise exists inevitably in the images,which results in that the number of intersection points of the projected parallel lines is more than one.In the authentic images,the intersection points are near each other.To insert one object into an image,the property of the vanishing point may be violated.The proposed approach makes minimal assumption,and the only requirement is that there is segment contained in the inserted object paralleled to the segments exist in the scene.The normalized mean value and the normalized standard deviation of the distances between the intersection points of the projected parallel lines are defined to be used as the evidence for separating authentic and tampered images.Secondly,a passive forensics approach based on the image measurement is proposed to solve the problem when the inserted object is with irregular shape can not be fixed by the above-mentioned method.In the image,the height ratio between the objects in the scene are distorted due to the perspective effect,which is molded by a pinhole camera.To insert one object into an image,the forger may not adjust the height of the object properly to satisfy the scene constraints.The proposed approach can determine the height ratio of two objects in an image without any knowledge of camera parameters.That is,the height ratio is found merely by the planar homography of the reference plane on which both objects of interest is situated and the vanishing point of the direction perpendicular to the reference plane.If the relative size relationship is not appropriate,then a forged image is identified.Experiments results prove the efficiency of the two proposed methods,and the detection accuracy of both methods is up to 100%.Compared with trace-based methods,our methods are less sensitive to the image operations that do not alter the content of images,such as lossy compression and image resampling.Up to present,scene perspective-based image composites detection is still in its preliminary exploration stage.It is hoped that the present study maybe promote the development of passive image forensics.